(681a) Accurate Property Prediction for Inorganic Materials with Machine Learning

Historically, novel materials have been discovered because of long and laborious trial-and-error process. Over the years, materials research has led to the accumulation of relatively large collections of experimental data on materials structure and properties. Prompted by the growth of materials databases, the emerging materials informatics approaches offer an opportunity to transform this trial-and-error practice into data- and knowledge-driven rational design and accelerated discovery of novel materials with the desired properties.

Using the data from the AFLOWLIB repository (http:www.aflowlib.org) of materials properties obtained with the high-throughput DFT calculations, we have constructed Machine Learning (ML) models to predict three critical materials properties including band gap, Fermi level energy, and the class of materials as metals or insulators. To enable these calculation, we have developed novel materials descriptors such as universal property-labelled fragments (PLMF).[1] We have established that the accuracy of predictions obtained with Quantitative Materials Structure–Property Relationship (QMSPR) models approaches that of GGA DFT functionals yet model development requires a minute fraction of computational time as compared to ab initio calculations. Notably, due to the representation of materials with PLMF the QMSPR models are broadly applicable to virtually any stoichiometric inorganic materials. This representation also affords straightforward model interpretation in terms of simple heuristic design rules that could guide rational design of novel materials. This proof-of-concept study demonstrates the power of materials informatics to dramatically accelerate the search for novel materials.

[1] O. Isayev, C. Oses, S. Curtarolo, A. Tropsha. The materials genome of electronic structure: prediction of band gap with machine learning. In press